The Silent Curriculum: How Does LLM Monoculture Shape Educational Content and Its Accessibility?

Supriti Vijay

​

Digital Marketing,

Adobe

Aman Priyanshu

​

School of Computer Science,

Carnegie Mellon University

Introduction

• Large Language Models (LLMs) are becoming the primary source of knowledge
• LLMs may propagate a singular perspective, creating an "LLM Monoculture"
• The "Silent Curriculum" shapes children's learning through LLM responses

Experimental Setup

• Utilized GPT-3.5 and LLaMA2-70B as subjects
• Generated an Ethnicity and Top 20 Occupations corpus
• LLMs created short stories about children's success in given occupations

Generating the Occupational-Racial Bias Benchmark

• LLMs generated occupations for each ethnic group
• Examples: White - "Corporate Executive", Black - "Music Producer", Asian - "Software Engineer"
• Occupations reflect cultural and stereotypical biases

​

We present our prompt for generation of these biased race-occupation pairs in Appendix A.1

Children's Story Generation

• LLMs crafted narratives about a child's journey to success in a specific occupation
• Prompts excluded direct mentions of ethnicity, focusing on occupation
• Aimed to probe implicit biases in the models' storytelling

Self-Annotating Ethnic/Racial Groups

• LLMs annotated ethnicities of characters in the generated stories
• Investigated self-consistency of models and disparities between biases and narratives
• Calculated cosine similarity scores to quantify consistency in cultural representation (Reference: Appendix A.3)

Results and Discussion

Comparison of LLM-specified occupational ethnicity counts against the inferred ethnicity of the protagonist's name. The heatmap illustrates discrepancies in portrayals, revealing potential biases in cultural representations within AI-generated narratives.

GPT3.5

LLaMA-70B

Heatmap illustrating discrepancies between LLM-specified occupational ethnicity counts and the inferred country. Provides insights into potential biases in cultural depictions within AI-generated content.

GPT3.5

LLaMA-70B

Questions We Need to Ask (Provocation)

1. The generation of biased occupation-racial/ethnic pairs: Is it safe?
1. LLMs generating stereotypical associations between occupations and ethnicities
2. Potential to reinforce harmful biases and limit children's aspirations

​

​

• Representation disparity in children's stories generated by LLMs
• Certain races more likely to occur, even when LLMs associate differently
• Are safety fine-tuning techniques biasing models to be "fair" according to inconsiderate benchmarks?

​

​

• High cosine similarity (0.86 and 0.87) across these two LLMs, at least, suggests bias convergence
• Likely due to the use of same/similar datasets and pre-training methods
• What are the implications of a homogenized AI perspective on society?

Conclusion

• LLMs outputs are influenced by societal biases and cultural narratives
• The "Silent Curriculum" may shape children's learning and perpetuate stereotypes
• A collective effort is needed to challenge and broaden the AI monoculture